What problem does it solve? Recovering prior work, commands, or decisions from a large local Codex history is difficult because session data is spread across session_index.jsonl, history.jsonl, and thousands of rollout JSONL files under active and archived roots. This Skill locates the smallest relevant transcript set and extracts only the decisive evidence without dumping entire files into context. ## Core Features & Use Cases - Session Recovery: Map a session ID, thread ID, date window, repo path, or user phrasing to exact rollout files across both active and archived session roots. - Bounded Scanning: Use scan_rollout.py for value-free structural orientation and field-aware literal search with strict byte, record, and output budgets. - Corpus Audits: Use build_session_corpus.py to build a deduplicated cross-root corpus for date-window audits, replay-prefix detection, and repeated workflow-friction analysis. - PR Attribution: Use pr_attribution.py to derive the exact OpenAI Codex PR attribution sentence from a task family's model and effort votes. - Use Case: When asked to "read your rollout" or audit skill trigger misses over the last week, inventory both transcript roots, select candidate rollouts, and scan each one with bounded field-aware search instead of raw grep. ## Quick Start Ask the assistant to use the codex-session-mining skill to find what was done in a recent Codex session by date or session ID.
